高级检索+

基于FOD-CNN光谱指数的滨海地区盐渍化估算模型

Estimating salinization in coastal areas using FOD-CNN spectral indices

  • 摘要: 为有效开展黄河三角洲地区盐渍化遥感监测,减少高光谱数据冗余,提高模型构建精度,该研究提出一种基于光谱指数的分数阶微分-卷积神经网络(FOD-CNN,fractional-order differential-convolutional neural networks)监测模型。该研究以东营市为研究区,对高光谱数据进行预处理及FOD变换处理,构建二维光谱指数进行敏感参量的筛选,采用卷积神经网络(CNN, convolutional neural networks)、偏最小二乘回归(PLSR,partial least squares regression)和随机森林(RF,random forest)模型开展对土壤含盐量监测的研究。分数阶微分变换能够有效突出光谱曲线变化过程中的渐变信息,且能够显著(P < 0.01)提高不同光谱指数与土壤含盐量之间的相关性;相较于PLSR模型和RF模型,CNN模型测试集相对分析误差(RPD,relative percent difference)分别提高1.74、1.76,决定系数(R2, coefficient of determination)分别提高0.03、0.28,均方根误差(RMSE,root mean square error)分别减少1.47、1.52 g/kg;CNN模型对轻、中、重及极重度盐渍化均表现出较好的反演效果,PLSR模型对极重度盐渍化反演效果较好,而RF模型反演效果均较差。该研究通过FOD与CNN模型相结合,提出了更适用于该地区盐分反演的更具有鲁棒性的FOD-CNN模型,可为研究黄河三角洲盐渍化的监测提供一定的技术支撑。

     

    Abstract: Soil salinization is relatively severe in the Yellow River Delta region. However, the conventional salinization monitoring can often suffer from the time-consuming, labor-intensive, and costly problems. Among them, the single-band information can be extracted from the hyperspectral data, in order to conduct the salinization monitoring. The inter-band correlation information cannot be fully utilized during extraction. Moreover, the original data can also share the low sensitivity to the salt response. This study aims to fully exploit the correlation information between spectra, particularly for the high sensitivity of spectral data. The accuracy of the salinity inversion was also improved to accurately determine the degree of salinization. Hyperspectral data source was collected from the field sampling in the Yellow River Delta region, Dongying City, Shandong Province, China. Pretreatment was also applied, including Savitzky-Golay (S-G) filtering and Multiplicative Scatter Correction (MSC). The sample data was split into the training and testing datasets at a 7:3 ratio. Fractional-order differentiation (FOD) processing was performed on the spectral data. Furthermore, nine typical indices of the two-dimensional spectra were constructed at each differential order, according to the possible pairwise combinations of the bands. The optimal band and differential combinations were obtained to evaluate the correlation coefficient between these indices and soil salt content. These feature variables were then selected to construct the various models, including Convolutional Neural Networks (CNN), Partial Least Squares Regression (PLSR), and Random Forest (RF). Grid search algorithm and ten-fold cross-validation were used for the hyperparameter tuning. Finally, the accuracy of the improved model was assessed using root mean square error (RMSE), relative percent difference (RPD), and the coefficient of determination (R2). Results showed that the noise and baseline drift were reduced in the preprocessed spectral curves. The smoother and more concentrated curves of the spectral features shared no significant change in the trend, compared with the original. The dataset all exhibited the strong variability, indicating the reasonable sample partitioning. FOD was effectively highlighted the gradual change during spectral curve variations, in order to enhance the sensitivity of the spectral data. The soil salinity was better predicted in the coastal saline areas. There was the significant correlation between different spectral indices and soil salt content (P<0.01). Compared with the original spectra, there were the significantly higher correlation coefficients between FOD-constructed spectral indices and soil salt content (P<0.01). The FOD with the spectral index was effectively identified the sensitive spectral information for the data dimensionality reduction. The Normalized Difference Index (NDI) at the 1.6 order (1244, and 2081 nm) also exhibited the highest correlation coefficient (0.9) with the soil salt content, followed by the Ratio Index (RI) at the 1.6 order (2242, 1208 nm) with a correlation coefficient of 0.88. The CNN model was achieved in the highest inversion accuracy, with the testing set RPD, R2, and RMSE values of 3.41, 0.91, and 1.42 g/kg, respectively. Compared with the PLSR and RF models, the CNN model's testing set RPD increased by 1.74 and 1.76, respectively, R2 increased by 0.03 and 0.28, respectively, whereas, RMSE decreased by 1.47 g/kg and 1.52 g/kg, respectively. The better inversion performance of the CNN model was achieved in the slight, moderate, severe, and very severe salinization. The PLSR model performed better for the very severe salinization, while the RF model performed poorly overall. The FOD was combined with the spectral indices, indicating the spectral sensitivity and inversion accuracy. Furthermore, the better adaptability of the CNN model was obtained after optimization, compared with the PLSR and RF models. A more robust FOD-CNN model was better suited for the salinity inversion, providing for the potential support to the monitoring salinization in the Yellow River Delta.

     

/

返回文章
返回